Fast-DENSER is a new extension to Deep Evolutionary Network Structured Evolution (DENSER). The vast majority of NeuroEvolution methods that optimise Deep Artificial Neural Networks (DANNs) only evaluate the candidate solutions for a fixed amount of epochs; this makes it difficult to effectively assess the learning strategy, and requires the best generated network to be further trained after evolution. Fast-DENSER enables the training time of the candidate solutions to grow continuously as necessary, i.e., in the initial generations the candidate solutions are trained for shorter times, and as generations proceed it is expected that longer training cycles enable better performances. Consequently, the models discovered by Fast-DENSER are fully-trained DANNs, and are ready for deployment after evolution, without the need for further training.
@article{assunccao2019fast,
title={Fast-DENSER++: Evolving Fully-Trained Deep Artificial Neural Networks},
author={Assun{\c{c}}{\~a}o, Filipe and Louren{\c{c}}o, Nuno and Machado, Penousal and Ribeiro, Bernardete},
journal={arXiv preprint arXiv:1905.02969},
year={2019}
}
@article{assunccao2018denser,
title={DENSER: deep evolutionary network structured representation},
author={Assun{\c{c}}ao, Filipe and Louren{\c{c}}o, Nuno and Machado, Penousal and Ribeiro, Bernardete},
journal={Genetic Programming and Evolvable Machines},
pages={1--31},
year={2018},
publisher={Springer}
}
CUDA >= 10; CuDNN>=7.0. Python3.7 or higher is required. The following python libraries are required: tensorflow, keras, numpy, sklearn, scipy, jsmin, and Pillow.
The datasets are located in the folder f-denser/fast_denser/utilities/datasets/data. In particular, we made available for download data for the svhn and tiny-imagenet datasets. The datasets are obtained from http://ufldl.stanford.edu/housenumbers/ and https://tiny-imagenet.herokuapp.com. To download the datasets simply execute the sh scripts:
sh tiny-imagenet-200.sh
sh svhn/svhn.sh
To install Fast-DENSER as a python library the following steps should be performed:
pip install -r requirements.txt
python setup.py install
python -m fast_denser.engine -d <dataset> -c <config> -r <run> -g <grammar>
-d [mandatory] can assume one of the following values: mnist, fashion-mnist, svhn, cifar10, cifar100-fine, cifar100-coarse, tiny-imagenet
-c [mandatory] is the path to a json configuration file. Check example/config.json for an example
-r [optional] the the run to be performed [0-14]
-g [mandatory] path to the grammar file to be used. Check example/modules.grammar for an example
The configuration file is formated in JSON. An example can be found here. A description of the parameters contained in the configuration file is below.
| Parameter Name | Description |
|---|---|
| random_seeds | Seeds for setting the initial random seeds for the random library. |
| numpy_seeds | Seeds for setting the initial random seeds for the numpy library. |
| lambda | Number of offspring to generate in each generation. |
| max_epochs | Maximum number of epochs to perform. Evolution is halted when the current number of epochs surpasses this value. |
| save_path | Place where the experiments are saved. |
| add_layer | Probability to add a layer to an individual (mutation) [0,1]. |
| reuse_layer | Probability to reuse a layer, when a new layer is added to an individual (mutation) [0,1]. |
| remove_layer | Probability to remove a layer from an individual (mutation) [0,1]. |
| add_connection | Probability to add a connection to the input of a layer (mutation) [0,1]. |
| add_connection | Probability to add a connection to the input of a layer (mutation) [0,1]. |
| remove_connection | Probability to remove a connection to the input of a layer (mutation) [0,1]. |
| dsge_layer | Probability to change any of the DSGE parameters, i.e., grammar expansion possibilities (mutation) [0,1]. |
| macro_layer | Probability to change any of the parameters of the macro-blocks (e.g., learning) (mutation) [0,1]. |
| train_longer | Probability to train a network for longer (mutation) [0,1]. |
| network_structure | Network structure, i.e., allowed sequence of layers. |
| output | Grammar production rule to use as the output layer. |
| macro_structure | Production rules to use as macro structure evolutionary units. |
| network_structure_init | Number of evolutionary units on initialisation. |
| levels_back | Number of levels back for each of the blocks. Settings values higher than one enables skip connections. |
| datagen | Data augmentation generator for the training data - keras interpretable. |
| datagen_test | Data augmentation generator for the test data - keras interpretable. |
| default_train_time | Maximum training time for each network (in seconds). |
| fitness_metric | Fitness assignment metric. |
You can also import Fast-DENSER as a usual python library. An example of the search of CNNs for the fashion-mnist dataset is presented next.
import fast_denser
fast_denser.search(0, 'fashion-mnist', 'example/config.json', 'example/cnn.grammar')The first parameter specifies the run, the second the dataset, the third the configuration file, and the last the grammar.
The units tests can be found in the f-denser/tests folder, and can be executed in the following way:
python3.7 -m tests.test_utils
python3.7 -m tests.test_grammar
The example seeks for Convolutional Neural Networks (CNNs) for the classification of the Fashion-MNIST dataset.
python3.7 -m fast_denser.engine -d fashion-mnist -c example/config.cfg -g example/cnn.grammar
CPU and GPU docker images are available at https://hub.docker.com/r/fillassuncao/f-denser.
The mapping procedure of the available codebase supports production rules that can encode either topology or learning evolutionary units. The layers must start by "layer:layer_type" where layer_type indicates the type of the layer, e.g., conv (for convolutional), or fc (for fully-connected). To the moment the available layer types are convolutional (conv), pooling (pool-max or pool-avg), fully-connected (fc), dropout (dropout), and batch-normalization (batch-norm). The learning production rules must start by "learning:algorithm", where the algorithm can be gradient-descent, adam, or rmsprop. An example of a grammar can be found in example/cnn.grammar.
The parameters are encoded in the production rules using the following format: [parameter-name, parameter-type, num-values, min-value, max-value], where the parameter-type can be integer or float; closed choice parameters are encoded using grammatical derivations. For each layer type the following parameters need to be defined:
| Layer Type | Parameters |
|---|---|
| Convolution | Number of filters (num-filters), shape of the filters (filter-shape), stride, padding, activation function (act), bias |
| Pooling | Kernel size (kernel-size), stride, padding |
| Fully-Connected | Number of units (num-units), activation function (act), bias |
| Dropout | Rate |
| Batch-Normalization | - |
For the learning algorithms the follow parameters need to be defined:
| Learning Algorithm | Parameters |
|---|---|
| Gradient-descent | Learning rate (lr), momentum, lr decay (decay), nesterov, batch size (batch_size), number of epochs (epochs), early stopping (early_stop) |
| Adam | Learning rate (lr), beta1, beta2, lr decay (decay), batch size (batch_size), number of epochs (epochs), early stopping (early_stop) |
| RMSProp | Learning rate (lr), rho, lr decay (decay), batch size (batch_size), number of epochs (epochs), early stopping (early_stop) |
The current grammar example focuses on the simultaneous optimisation of the topoogy and learning strategy. In case the user only intends to optimise the topology, the learning can be fixed by replacing the learning production rule by for example: " <learning> ::= learning:gradient-descent lr:0.01 momentum:0.9 decay:0.0001 nesterov:True". The same rationale applies to the topology.
The required parameters, and layers can be easily changed / extended by adapting the function that performs the mapping from the phenotype into a keras interpretable model. See the next section for further details.
To add new layers (or simply change the mandatory parameters) one needs to add (or adapt) the mapping from the phenotype to the keras interpretable model. This can be easily performed by adding the necessary code to the utils.py file, in the "assemble_network" function of the Evaluator class (starting in line 244). The code is to be added between the "#Create layers -- ADD NEW LAYERS HERE" and "#END ADD NEW LAYERS" comments. To change the parameters of an already existing layer there is just the need to change the call to the keras layer constructor. To add new layers a keras layer constructor must be added, and the parameters passed to it. For example, to add a Depthwise Seperable 2D Convolution we would write the following code:
elif layer_type == 'sep-conv':
sep_conv = keras.layers.SeparableConv2D(filters = int(layer_params['num-filters'][0]),
kernel_size = (int(layer_params['kernel-size'][0]), int(layer_params['kernel-size'][0])),
strides = (int(layer_params['stride'][0]), int(layer_params['stride'][0])),
padding = padding=layer_params['padding'][0],
dilation_rate = (int(layer_params['dilation-rate'][0]), int(layer_params['dilation-rate'][0])),
activation = layer_params['act'][0],
use_bias = eval(layer_params['bias'][0]))
layers.append(sep_conv)In addition, to enable the use of the above layers in evolution, we would need to add a new production rule to the grammar: "<separable-conv> ::= layer:sep-conv [num-filters,int,1,32,256] [kernel-size,int,1,2,5] [stride,int,1,1,3] <padding> [dilation-rate,int,1,1,3] <activation-function> <bias>"
The addition of new fitness functions follows the rationale of the addition of new layers. We need to create the necessary code and add it to the fitness_metrics.py file. Currently it supports the accuracy, and the mean squared error. For example, to add the root mean squared error we can add the following code:
def rmse(y_true, y_pred):
from math import sqrt
return sqrt(mse(y_true, y_pred))After adding the rmse function we can use the rmse in the config.json file.
Any questions, comments or suggestion should be directed to Filipe Assunção ([email protected])